AMD Vega Hardware Reviews

[CarstenS]

PCPer's preview isn't without its fair share of twilight zone material, like the Fury X consuming 180W in some benchmarks.

That Tomb Raider benchmark? Bottlenecking from tessellation saves a lot of power. Conveniently, that seems also the test were Fury X is furthest behind Vega.

Why is the Fury X so efficient?

Efficiency always takes performance into account. How is Fury X's performance in this particular test?

 

[kalelovil]

GCN4 achieves ~8% higher IPC than GCN3, so to see Vega at or behind GCN3 IPC is certainly odd.
http://www.hardware.fr/getgraphimg.php?id=426&n=1

The only obvious thing would be a memory bottleneck, if tiling-based binning and improved delta-color compression hasn't achieved the efficiency AMD expected.

 

[silent_guy]

As much as memory BW is important, it can't explain everything. A 1070 has 256GB/s. FE has almost 500.

 

[yuri]

The only obvious thing would be a memory bottleneck

How about uarch changes aimed at reaching higher freq? A layman would guess eg. latent caches.

 

[ArkeoTP]

As much as memory BW is important, it can't explain everything. A 1070 has 256GB/s. FE has almost 500.

I really hope not but it's possible they messed up the IMC again and the card's effective memory bandwidth is far below the theoretical bandwidth. I don't think it's enough to explain the regression though.

 

[Clukos]

https://twitter.com/x/status/881231445276184576

https://twitter.com/x/status/881357928451772418

 

[Clukos]

http://www.gamersnexus.net/hwreviews/2973-amd-vega-frontier-edition-reviewed-too-soon-to-call

 

[rcf]

Gamersnexus puts Vega FE between gtx1070 and gtx1080 in gaming performance, but for some strange reason they removed both the 1070 and 1080 from the final power consumption results, leaving only Titan/1080Ti results which are higher and closer to Vega FE's.
They should have also included at least one Quadro card in the professional benchmarks.

 

[Rootax]

They removed the Fury X in game bench too, with is sad because in some game it's on par with 1070. Being a Fury X owner, my interest is to see what gain exists between the two...

 

[ArkeoTP]

PCPer updated their "doe" size measurement to 512mm^2

https://twitter.com/x/status/881317151277285376

There's also this tweet by Raja Koduri, which implies a die size of 484mm^2 (22*22)

https://twitter.com/x/status/881370308694822913

So ummm... Yeah...

 

[CarstenS]

512 including bonding residues and all other inaccuracies from manual measurement is close enough for me to the promised 500. Awkward that PCPer was so far off. Oh the calamities of a time-pressured review.

 

[Alexko]

PCPer updated their "doe" size measurement to 512mm^2

https://twitter.com/x/status/881317151277285376

There's also this tweet by Raja Koduri, which implies a die size of 484mm^2 (22*22)

https://twitter.com/x/status/881370308694822913

So ummm... Yeah...

That would be 484 mm², then—22 mm × 22 mm.

 

[Nemo]

PCPer updated: 564mm2 >> 512mm2 >> .... To be continued.

 

[Deleted member 2197]

Gamersnexus puts Vega FE between gtx1070 and gtx1080 in gaming performance, but for some strange reason they removed both the 1070 and 1080 from the final power consumption results, leaving only Titan/1080Ti results which are higher and closer to Vega FE's.
They should have also included at least one Quadro card in the professional benchmarks.

Looks like they also removed the Vega FE power/thermal "overclock results", but display the TitanXp results in the System power/consumption bar graphs.

 

[CarstenS]

Why would they do that?

 

[Deleted member 2197]

Why shouldn't they do that if they are attempting to make comparisons? I'd assume it was tested similar to the TitanXp.

 

[CarstenS]

I mean, why would they remove results? Or weren't those results there in the first place and they only omitted them?

 

[Deleted member 2197]

Not sure. Looking at the comments someone reminded him(3 hrs ago) to include "For Honor" results, so maybe they will add missing pieces throughout the day.

 

[xEx]

Its also funny how when you try to OC the card the HBM clocks drops to half.

 

[kalelovil]

I mean, why would they remove results? Or weren't those results there in the first place and they only omitted them?

WattMan bug resulting in memory frequency being halved after core is overclocked. Thus their 'OC' results showed lower performance than stock, and are not a setting an end-user would ever want.